Abstract Book

S183

ESTRO 37

scored based on RTOG toxicity scale. Biochemical relapse was defined as more than 2 ng/ml increase after the PSA nadir had been reached. The two treatment arms were compared with the Mann-Whitney U test. Results: The median follow up time was 23 months (5-34 months). Four patients had biochemical relapse, two in the LDR (4%) and two (4%) in the HDR BT arm. Three biochemical failures were detected in the IR, and one in the LR group. In all patients with biochemical relapse a local failure was confirmed. No regional or distant metastases were detected. Acute ≥grade2 gastrointestinal (GI) toxicity occurred only in one patient in the LDR arm. Acute ≥ grade2 urogenital (UG) toxicity occurred in 13 patients (26%) in the HDR and in 36 patients (72%) in the LDR group (p<0,001). Acute grade 3 toxicity was only in 1 (2%) patient in the HDR and in two (4%) patients in the LDR group. The mean IPSS score 3 months after BT was significantly better with HDR than with LDR BT (8.9 vs 15.9, p=0,0004), although this difference disappeared in 12 months after BT (8.3 vs 6.9, p=0.3). On the last follow up no ≥ grade2 GI toxicity could be detected, grade 2 UG side effects remained in 7 patients (14%) in the HDR and in 13 patients (26%) in the LDR group. Late grade 3 UG toxicity were detected only in one (2%) patient in the HDR and in two patients (4%) in the LDR group. The V100 was significantly better with LDR than HDR (98.8% vs 97.3%, p<0,001), meanwhile the values of Du10% and Du30% for the urethra were better with HDR compared to LDR technique (114% vs 132%, p<0,001 and 111% vs 128%, p<0,001). Conclusions: In our randomized study there were no differences in biochemical control or local failure rate in patients treated with LDR or single fraction HDR monotherapy for localized prostate cancer. Acute urological side effects were significantly worse with LDR BT, but the difference disappeared by 1 year after BT. More follow up is needed to confirm our results.

Application of machine learning currently covers varying areas of the radiation oncology field ranging from treatment response modelling, treatment planning, image-guidance, motion tracking, and quality assurance in radiation oncology. In this talk, we will briefly review some of the basic principles of machine learning algorithms and provide representative examples of their current application in radiation oncology. We will discuss their advantages and how to avoid some of the common pitfalls (the Do’s and Don’ts of machine learning) with special focus on overfitting issues and best practice strategies. We will highlight some of the current challenges for their application in radiation oncology and potential avenues for future research. SP-0354 Machine learning for image segmentation D. Rueckert 1 1 Imperial College London, Department of Computing, LONDON, United Kingdom Abstract text This talk will introduce the concept of machine learning for medical image segmentation. We will review different deep learning approaches for this task, focusing on different network architectures such as fully connected networks (FCN), U-Net and DeepMedic. We will review the advantages and disadvantages of these different architectures in the context of segmentation in neuroimaging, cardiac imaging and cancer imaging. As these various networks architectures perform differently in real-world applications, with behaviour largely influenced by architectural choices and training settings, we will also explore the concept of Ensembles of Multiple Models and Architectures (EMMA) which enables robust performance through aggregation of predictions from a wide range of methods. The approach reduces the influence of the meta-parameters of individual models and the risk of overfitting the configuration to a particular database. EMMA can be seen as an unbiased, generic deep learning model which has been shown to yield excellent performance in recent segmentation challenges. Finally, we will review unsupervised domain adaptation using adversarial neural networks to train segmentation models which is more invariant to differences in the input data, and which do not require any annotations on the test domain. SP-0355 Machine learning for radiomics and outcome modeling M. Hatt 1 1 INSERM UMR 1101 - LaTIM, Department of Radiation Oncology, Brest, France Abstract text Radiomics has seen an exponential growth as a field of research in cancer outcome modeling. The high- throughput extraction of quantitative data from multimodal medical images opens the way to more discriminative prognostic and predictive models allowing for better patient management. However, numerous challenges still have to be addressed before radiomics- derived multiparametric models can actually be used in clinical practice, including but not limited to: standardization of image analysis and features computation workflow for better reproducibility of studies, machine learning algorithms choice and optimization and multi-centric validation of the developed models. In this talk, a critical review of recently proposed machine learning methods (including deep learning developments) for radiomics will be proposed, as well as insights into the potential future challenges to address

Debate: The technological advancement in radiotherapy in the last decade was driven by economics, not clinical needs

SP-0351 For the motion (advancement driven by economics) C. Kirisits Medical University of Vienna, Vienna, Austria

Abstract not received

SP-0352 Against the motion M. Van Vulpen The Netherlands

Abstract not received

Symposium: Applications of machine learning in radiation oncology

SP-0353 Introduction to Machine Learning and its Application in Radiotherapy I. El Naqa 1 1 University of Michigan, Radiation Oncology, Ann Arbor, USA Abstract text With the era of big data and the rapid increase in patient-specific information, there has been burgeoning interest by the radiation oncology community to utilize machine learning (ML) algorithms because of their ability to learn data dependencies from the current environment and generalize into unseen tasks.